International Journal of Applied Earth Observation and Geoinformation 22 (2013) 40–52

Contents lists available at SciVerse ScienceDirect

International Journal of Applied Earth Observation and

Geoinformation
journal homepage: www.elsevier.com/locate/jag

Night on Earth: Mapping decadal changes of anthropogenic night light in Asia

Christopher Small a,∗ , Christopher D. Elvidge b
a
Lamont Doherty Earth Observatory, Columbia University, New York City, USA
b
Earth Observation Group, NOAA National Geophysical Data Center, Boulder, CO, USA

a r t i c l e i n f o a b s t r a c t

Article history: The defense meteorological satellite program (DMSP) operational linescan system (OLS) sensors have
Received 8 July 2011 imaged emitted light from Earth’s surface since the 1970s. Temporal overlap in the missions of 5 OLS
Accepted 24 February 2012 sensors allows for intercalibration of the annual composites over the past 19 years (Elvidge et al., 2009).
The resulting image time series captures a spatiotemporal signature of the growth and evolution of
Keywords: lighted human settlements and development. We use empirical orthogonal function (EOF) analysis and
Urban
the temporal feature space to characterize and quantify patterns of temporal change in stable night light
Night light
brightness and spatial extent since 1992. Temporal EOF analysis provides a statistical basis for represent-
DMSP-OLS
Landsat
ing spatially abundant temporal patterns in the image time series as uncorrelated vectors of brightness
Zipf as a function of time from 1992 to 2009. The variance partition of the eigenvalue spectrum combined
Asia with temporal structure of the EOFs and spatial structure of the PCs provides a basis for distinguishing
India between deterministic multi-year trends and stochastic year-to-year variance. The low order EOFs and
China principal components (PC) space together discriminate both earlier (1990s) and later (2000s) increases
Nightsat and decreases in brightness. Inverse transformation of these low order dimensions reduces stochastic
variance sufficiently so that tri-temporal composites depict potentially deterministic decadal trends.
The most pronounced changes occur in Asia. At critical brightness threshold we find an 18% increase
in the number of spatially distinct lights and an 80% increase in lighted area in southern and eastern
Asia between 1992 and 2009. During this time both China and India experienced a ∼20% increase in
number of lights and a ∼270% increase in lighted area – although the timing of the increase is later in
China than in India. Throughout Asia a variety of different patterns of brightness increase are apparent
in tri-temporal brightness composites – as well as some conspicuous areas of apparently decreasing
background luminance and, in many places, intermittent light suggesting development of infrastruc-
ture rather than persistently lighted development. Vicarious validation using higher resolution Landsat
imagery verifies multiple phases of urban growth in several cities as well as the consistent presence of low
DN (<∼15) background luminance for many agricultural areas. Lights also allow us to quantify changes
in the size distribution and connectedness of different intensities of development. Over a wide range of
brightnesses, the size distributions of spatially contiguous lighted area are consistent with power laws
with exponents near −1 as predicted by Zipf’s Law for cities. However, the larger lighted segments are
much larger than individual cities; they correspond to vast spatial networks of contiguous development
(Small et al., 2011).
© 2012 Published by Elsevier B.V.

1. Introduction intercalibration of the 30 annual composites over the past 19 years

The defense meteorological satellite program (DMSP) opera-
tional linescan system (OLS) has imaged emitted light from Earth’s
surface since the 1970s. In 1992 a digital archive was established
for DMSP-OLS data at the NOAA National Geophysical Data Cen-
ter. Annual composites of temporally stable night light have been
processed and distributed by the NGDC for every year since 1992.
Temporal overlap in the missions of multiple OLS sensors allows for
∗ Corresponding author.
E-mail address: small@LDEO.columbia.edu (C. Small).
(Elvidge et al., 2009). These data offer the opportunity to quantify
a unique spatiotemporal signature of human settlement growth
and evolution. Here we summarize some initial results of a spa-
tiotemporal analysis of annual change in the extent, distribution
and intensity of anthropogenic night light in Asia.
The objective of this study is to quantify apparent changes in
the number, extent and brightness of stable night lights in Asia.
We limit the geographic scope of this study to Asia because this is
where we observe the greatest changes. The intention is to illus-
trate a methodology for characterization and vicarious validation
of apparent changes in areas where anthropogenic changes are
large enough to be clearly distinguished from other types of change

0303-2434/$ – see front matter © 2012 Published by Elsevier B.V.

doi:10.1016/j.jag.2012.02.009
 C. Small, C.D. Elvidge / International Journal of Applied Earth Observation and Geoinformation 22 (2013) 40–52
related to sensor calibration and atmospheric effects. The first part
of the study illustrates the use of the principal component (PC)
transformation and empirical orthogonal function (EOF) analysis
with multi-temporal night light data. Background and details of the
EOF analysis are given by Small (under review). Here we present the
results of the EOF analysis and discuss the spatiotemporal changes
quantified by the analysis. The second part of the study presents
a vicarious validation of changes in night lights using changes in
land cover as imaged by Landsats 5 and 7. The third part of the
study presents spatial and temporal variations in the area, num-
ber and rank size distributions of lighted areas and discusses the
implications for urban growth processes.
2. Mapping night light from space
Since 1994, NGDC has had an active program focused on
global mapping of nighttime lights using the data collected by
the DMSP-OLS sensors. The basic product is a global annual
cloud-free composite, which averages the OLS visible band data
for one satellite from the cloud-free segments of individual
orbits. By compositing only the cloud-free segments the impact
of atmospheric effects is minimized. Over the years, NGDC has
developed automatic algorithms for screening the quality of the
nighttime visible band observations to remove areas contaminated
by sunlight, moonlight, and the presence of clouds (Baugh et al.,
2009; Elvidge et al., 1997). For this study we used the NGDC
annual stable lights product. In the stable lights annual cloud-free
composite product generation, fires and other ephemeral lights
are removed based on their high brightness and short duration.
Background noise is removed by setting thresholds based on
visible band values found in areas known to be free of detectable
lights. In 2010, NGDC released the version 4 time series of Stable
Lights, spanning the years 1992–2009. These are available online at
http://www.ngdc.noaa.gov/dmsp/downloadV4composites.html.
Because the OLS visible band has no on-board calibration system,
the data from each satellite year should be intercalibrated prior
to quantitative analysis of the time series. Baugh et al. (2009)
provides a description of the algorithms used to make the v.4
stable lights products. The intercalibration procedure is described
by Elvidge et al. (2009). This intercalibration is applied to the time
series of annual composites used in our analysis.
Temporally stable night lights, as measured by the Defense
Meteorological Satellite Program-Operational Line Scanner (DMSP-
OLS), provide a unique proxy for anthropogenic development. In
previous studies brightness and spatial extent of emitted light has
been correlated to population density (Briggs et al., 2007; Sutton
et al., 2001), built area density (Elvidge et al., 2007), economic activ-
ity (Doll et al., 2006; Ghosh et al., 2010; Henderson et al., 2009), and
electrification rates (Elvidge et al., submitted for publication). Night
lights are known to overestimate the spatial extent of development
at the periphery of settlements (Elvidge et al., 1997) but the spa-
tial extent of this overglow is generally small compared to the total
area of all but the smallest settlements (Small et al., 2005). Com-
parisons of stable light with 30 m resolution Landsat imagery on a
wide variety of population density gradients indicates that average
brightness increases with increasing spatial density of Shortwave
Infrared reflectance and shadow associated with constructed sur-
faces (Small et al., 2005) as well as actual maps of impervious
surfaces (Elvidge et al., 2007).
In this study, we use tri-temporal change maps to illustrate
decadal changes in brightness and spatial extent of night lights.
Displaying brightness of three annual average brightness compos-
ites in the red, green and blue channels of a color image highlights
areas of change as color deviations from the gray shades that indi-
cate equal brightnesses in all three years. A tri-temporal composite
41
of 1994, 2001 and 2009 for southeastern Asia is shown in Fig. 1. At
the scale of this figure some large area changes in background lumi-
nance are visible but most of the changes are too small to be seen.
A linear stretch between 0 and 12 DN highlights changes in back-
ground luminance at the expense of saturating changes in brighter
more developed areas. This figure illustrates two challenges of
representing change with tri-temporal brightness composites: (1)
showing change in both bright and dim areas simultaneously
and (2) showing deterministic decadal changes in the presence
of spurious year-to-year variability. In order to circumvent these
challenges, and make better use of the information content of the
composite time series, we present an approach that uses all 30
annual composites to quantify change in lighted areas. However,
we retain the convenience of using tri-temporal color compos-
ites to illustrate these changes. We accomplish this by separating
distinct components of spatiotemporal change using EOF analy-
sis to represent the interannual trends of the observed changes
in night light. As explained in detail below, the EOF analysis pro-
vides an estimate of the dimensionality of the image time series
and a basis for separating deterministic and stochastic compo-
nents of the spatiotemporal changes observed. In the context of
this study, dimensionality refers to the different background levels
and temporal patterns of change and their corresponding spatial
distributions. We use the EOF analysis to minimize the effect of
stochastic year-to-year variance in brightness thereby emphasiz-
ing interannual trends in the brightness and spatial extent of lighted
area. The structure of the temporal feature space (explained below)
also provides a concise representation of the overall distribution of
changes in the night light data.
3. Spatio-temporal dimensionality and empirical
orthogonal function analysis
Principal component transformations are commonly used to
represent uncorrelated modes of variance in high dimensional data.
Different types of PC transform are used to reduce the dimension-
ality of multispectral imagery (e.g. Green et al., 1988; Lee et al.,
1990; Singh and Harrison, 1985) and to represent the topology of
spectral feature spaces (Adams et al., 1986; Crist and Cicone, 1984;
Johnson et al., 1985; Kauth and Thomas, 1976; Smith et al., 1985).
Because spectral bands are often correlated, PC transforms provide
an efficient low dimensional projection of the uncorrelated com-
ponents of the spectral feature space. The same property applies
to temporal dimensions. PC transforms have also been used to rep-
resent uncorrelated patterns in multi-temporal imagery (Richards,
1984) (Eastman and Fulk, 1993; Townshend et al., 1985) and for
change detection (Byrne et al., 1980; Fung and LeDrew, 1987). In
meteorology and oceanography the PC transformation provides the
basis of empirical orthogonal function analysis; a standard tool for
analysis of spatio-temporal patterns and processes (see Bretherton
et al., 1992; Preisendorffer, 1988; von Storch and Zwiers, 1999 for
overviews).
The utility of the PC transform for representing spatiotempo-
ral processes is related to the fact that any location-specific (x)
pixel time series Pxt contained in an N image time series can
be represented as a combination of temporal patterns and their
location-specific components as:

N
Pxt = Cix Fit (1)
i=1
where Cix is the spatial principal component (PC) and Fit is the corre-
sponding temporal empirical orthogonal function (EOF) and i is the
dimension. The EOFs are the eigenvectors of the covariance matrix.
EOFs represent uncorrelated temporal patterns of variability within
the data. The PCs are the corresponding weights that represent the
42 C. Small, C.D. Elvidge / International Journal of Applied Earth Observation and Geoinformation 22 (2013) 40–52

Fig. 1. Tri-temporal night light brightness change map of eastern Asia derived from 3 of 30 annual composites of DMSP-OLS imagery compiled and intercalibrated by Elvidge
et al. (2009). Temporal overlap among the five OLS sensors shown inset allows for intercalibration among annual composites. Average annual brightness of cloud-free images
ranges from 0 to 64 DN but a linear stretch between 0 and 12 DN emphasizes change in smaller dimmer lights at the expense of saturation of brighter lights. Note the regional
scale consistencies in change of dimmer lights between large cities – particularly in India and Thailand. (For interpretation of the references to color in this figure text, the
reader is referred to the web version of the article.)

relative contribution of each EOF to the corresponding pixel time
series Pxt at each location x. The relative contribution of each EOF
to the total spatio-temporal variance is given by the eigenvalues of
the covariance matrix. The distribution of eigenvalues also gives an
indication of the dimensionality of the data in terms of uncorrelated
modes of variance.
In this study dimensionality refers to the structure of the spa-
tiotemporal patterns represented in the data – and their relative
magnitude compared to the stochastic variance. In oceanography
and meteorology the PC transformation is often used to charac-
terize dynamically important modes of spatiotemporal variance
associated with physical processes. The implicit assumption is that
some number, D (≤N), of the low order EOFs and their corre-
sponding PCs represent deterministic processes and that the higher
order dimensions represent stochastic variance ε. This allows an
observed pixel time series to be represented as a sum of determin-
istic and stochastic components as:

D
Pxt = Cix Fit + ε
i=1
EOFs are generally spatial patterns intended to represent spa-
tially continuous modes of variability of physical processes while
the PCs are generally the weights representing the temporal contri-
bution of the corresponding spatial pattern (Preisendorffer, 1988;
von Storch and Zwiers, 1999). In this study the convention is
reversed so the EOFs represent temporal patterns and the PCs
represent the corresponding spatial weights. In conventional EOF
analysis, there is often an attempt to interpret different EOFs in
terms of dynamical processes. However, this can be challenging
because the EOFs merely represent statistically uncorrelated modes
of variance but are not necessarily dynamically distinct. In the con-
text of this study we consider multiyear interannual to decadal
trends to result from deterministic processes and higher frequency
year-to-year variability to result primarily from stochastic pro-
cesses related to actual brightness variations as well as spatial and
temporal sampling effects. We acknowledge that some year-to-
year variability may actually result from deterministic processes
but that it may be effectively indistinguishable from stochastic
variability in these data.
In this analysis we show that the low order EOFs of night lights
do represent distinct components of change but not necessarily
independent processes. The diversity of processes responsible for
presumed deterministic changes in night light extent and bright-
ness can be represented in terms of both average brightness and
(2)
the sense and timing of change. The temporal patterns are given by
the EOFs while the relationships among these patterns are given by
the topology of the temporal feature space with dimensions defined
by the low order PCs. The dimensionality is indicated by the size
distribution of the eigenvalues. For this analysis we use the uncen-
tered, unscaled PC transform based on the eigenstructure of the
covariance matrix – but the method we present could also use the
 C. Small, C.D. Elvidge / International Journal of Applied Earth Observation and Geoinformation 22 (2013) 40–52 43

Fig. 2. Eigenstructure and temporal feature space for the Asia night light image time series. (a) The eigenvalues attribute ∼96% of variance to the first 3 dimensions. Higher
order eigenvalues diminish continuously. Correlation matrix (inset left) shows strong year to year correlation. Eigenvector (EOF) matrix (inset right) is dominated by high
frequency structure with only 3 relatively continuous EOFs. Inset plots of EOFs 1–4 show the temporal mean (EOF 1), the multi-decadal trend (EOF 2), the decadal trend (EOF
3) and a less obvious pattern (EOF 4). (b) The topology of the cloud in temporal feature space is dominated by the bimodal distribution of bright and dim lights displayed
in the first PC. The PC 2/3 projection represents contributions of the linear 18 year trend of EOF 2 and the half cycle of EOF 3. The diversity of decadal patterns results in a
constellation of distinct EOF 2/3 combinations surrounding the bimodal brightness axis in PC1. Inset numbers on end view correspond to time series in Fig. 3.

transform based on the correlation matrix with or without scal-
ing. Additional details of the approach are given by Small (under
review).
4. Characterizing spatio-temporal change
We use empirical orthogonal function (EOF) analysis to quan-
tify patterns of temporal change in OLS stable night light brightness
and spatial extent over the past 19 years. Intercalibrated data are
stacked into a spatial time series of images of annual average
brightness resulting from both brightly lit cities and from dimmer,
less intensively developed rural areas. Meaningful combinations
of EOFs can be derived from the topology of the temporal feature
space. The temporal feature space, defined by the low order princi-
pal components, represents the spatial distribution of image pixel
time series in terms of the relative contribution of the temporal
EOFs (Small, under review). The approach used here is based on
EOF analyses commonly used to represent geophysical processes
in oceanography and meteorology (Storch and Zwiers, 1999) – but
casts the EOFs as temporal rather than spatial modes and uses the
topology of the low order PC space to determine which linear com-
binations of low order EOFs depict physically meaningful temporal
patterns (Small, under review).
The spatio-temporal variance in the night light data is rep-
resented by the eigenstructure of the covariance matrix and the
topology of the temporal feature space. These are illustrated in
Fig. 2. The eigenvalues show that the first two dimensions account
for much more variance than the remaining continuum of dimen-
sions. The structure of the three low order EOFs is also more
continuous than the higher order EOFs. The topology of the tempo-
ral feature space shows the combination of EOFs that represent
most of the spatiotemporal variance in the image time series.
Fig. 2 shows the variance structure of the data and the relation-
ship between the spatial PCs and corresponding temporal EOFs
in the three low order PC dimensions that represent 91% of the
total variance. The separation of the overall brightness from the
decadal changes is apparent in the structure of the temporal fea-
ture space. The topology of the feature space is similar to that of
44 C. Small, C.D. Elvidge / International Journal of Applied Earth Observation and Geoinformation 22 (2013) 40–52

Fig. 3. Example time series from inner and outer radial peripheries of the PC3/2 projection of the temporal feature space in Fig. 2. Raw trajectories (red) are generally well
represented by 6 EOF (blue) and often by 3 EOF (green) projection filtering but more abrupt changes (e.g. 2o, 7o, 10i, 12o) are not captured by only 3 EOFs. Note progressive
shift of extrema with location in PC 3/2 space. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of the article.)

an HSV color space. In both spaces the brightness forms a central
axis about which a circular continuum of temporal changes (peak
years = hues) are distributed. This color analogy extends naturally
to the tri-temporal change map. When a single year of brightness
is represented by a primary color channel (R, G, or B) the temporal
changes in brightness correspond to continuous variations in hue
– both in the HSV analogy and in the composite image.
The intercalibration of the annual night light composites con-
siderably reduces spurious interannual variability – but does not
eliminate it entirely. Significant year to year variability remains in
the intercalibrated image time series. Much of this spurious vari-
ance is related to spatial uncertainty in the coregistration of the
annual composites. In addition, the phenomenon of low luminance
“overglow” results from a combination of limited sensor resolution,
atmospheric scattering and subannual spatial coregistration uncer-
tainty in the compositing process. This overglow causes spatial
blurring within individual annual composites. Additional tempo-
ral blurring results from interannual registration uncertainty. The
result is a combination of spurious spatial blurring and temporal
variance superimposed on whatever actual changes in brightness
 C. Small, C.D. Elvidge / International Journal of Applied Earth Observation and Geoinformation 22 (2013) 40–52
Fig. 4. Projection filtered tri-temporal night lights for Asia. Only EOFs 2 and 3 are back-projected so brightness variations are removed as well as higher order variance
unrelated to decadal brightness trends.
may be measured by the sensor. Spatial or temporal filtering has
the undesirable effect of mixing this spatiotemporal noise with the
signal in unknown proportions.
We address the spatial blurring and temporal noise problems
by using the information contained in the temporal EOFs and their
spatial abundance and distribution in the temporal feature space.
The projections of the low order PCs reveal the relationship among
the low order EOFs in terms of the actual temporal EMs that span
the space of all temporal patterns present in the image time series.
The variance partition of the PC transform makes use of the spatial
abundance of different temporal patterns to distinguish the com-
ponents of the temporal patterns that the greatest number of pixel
time series have in common. Two orthogonal projections of the
three low order PCs for the night light time series for southeast-
ern Asia are shown in Fig. 2 – along with the 3 low order EOFs
that represent ∼96% of the variance in the image time series. EOF 1
represents the temporal mean brightness of the image time series
and accounts for ∼91% of the variance. EOFs 2 and 3 represent the
decadal trends that explain the next 5% of variance. The remaining
27 modes each represent less than 1% and together less than 4% of
the total variance in the image time series. On the basis of the clear
break in the slope of the eigenvalue spectrum and the smoothness
of the 3 low order temporal EOFs, we infer that the interannual
growth component of the night light data is represented primar-
ily by the first three dimensions and attribute the remaining 4% of
variance in dimensions >3 to stochastic and local processes.
The structure of the projection of PCs 2 and 3 highlight the
difference between early (1990s) and late (2000s) increases and
decreases in the two decade time series. By inverse transform-
ing only dimensions 2 and 3 we can simultaneously remove both
the spurious variance and the dominant mean brightness from the
45
time series. This emphasizes only the decadal trends represented
by early and late increases and decreases in brightness. In the tem-
poral dimension, the result of this projection filtering is smoothed
time series lacking the spurious variance that is not reflected in
the low order EOFs depicting the most spatially abundant tempo-
ral patterns (Fig. 3). The projection filtered image time series does
not contain the spurious year-to-year variability and allows the
decadal trends of bright and dim lights to be shown simultaneously
on the same map. Fig. 4 shows a projection filtered tri-temporal
composite of the same years shown in Fig. 1. Note that the saturated
bright areas in Fig. 1 are replaced by primary colors emphasizing
the decadal trends.
Projection filtering depicts the decadal patterns of change in
night light brightness without the distraction of spurious year-to-
year differences in apparent brightness and the ten-fold difference
in brightness between the bright, heavily developed urban cen-
ters and the much dimmer peripheral areas where development
and urban growth begins. The tri-temporal change maps presented
here use the same display conventions to compare tri-temporal
composites of intercalibratted night light with a projection filtered
tri-temporal composite of the same annual composites with only
the decadal patterns associated with the 2nd and 3rd dimensions
of the PCs and EOFs. In both composites warmer colors (red, yel-
low) indicate brightening over the past 18 years. Cooler colors
(blue, cyan) indicate dimming. Green areas show some brighten-
ing during the 1990s followed by some dimming in the 2000s;
magenta areas have the opposite pattern. The effect of projec-
tion filtering on brightness time series is shown in Fig. 4. Inverse
transforming 3 low order dimensions represents most decadal
trends accurately but smooths some abrupt changes (e.g. 7o).
Inverse transforming 6 low order dimensions better represents
46 C. Small, C.D. Elvidge / International Journal of Applied Earth Observation and Geoinformation 22 (2013) 40–52

Fig. 5. Tri-temporal Landsat change comparison for New Delhi. Note infilling development, increase in shadow fraction (darkening) and radial road construction.

these abrupt changes while still eliminating spurious year-to-year
variance.
5. Comparisons with Landsat
Multitemporal Landsat imagery provides independent corrob-
oration of land cover changes suggested by observed changes
in night light. In areas with sufficient Landsat coverage we can
use changes in land surface reflectance to infer changes in land
cover that may correspond to changes in anthropogenic lighting
detected by OLS. Here we represent changes in exoatmospheric
reflectance in terms of spectral endmember fractions (Small,
2005). Global analyses of Landsat imagery show that ∼98% of
land surface reflectance can be represented accurately as linear
 C. Small, C.D. Elvidge / International Journal of Applied Earth Observation and Geoinformation 22 (2013) 40–52 47

Fig. 6. Tri-temporal Landsat change comparison for Kuala Lumpur. Increased brightening is disproportionately larger than developed land area as brightness of development
varies.

mixtures of rock–soil substrate, vegetation and dark sur-
face/shadow (Small, 2004). Global analyses of urban reflectance
show that most impervious surfaces are spectrally indistin-
guishable from rock–soil substrates when aggregated spatially
and spectrally to Landsat resolutions. In temperate and tropical
environments, where undeveloped land is generally vegetated,
high substrate fractions usually correspond to anthropogenic
development or fallow agricultural areas. To illustrate the cor-
respondence between spatiotemporal changes in land cover and
changes in night light we compare tri-temporal maps of Landsat –
derived substrate fraction to contemporaneous tri-temporal maps
of night light.
Tri-temporal night lights reveal a pronounced increase in
brightness extending southeastward from New Delhi since 2000
48 C. Small, C.D. Elvidge / International Journal of Applied Earth Observation and Geoinformation 22 (2013) 40–52

Fig. 7. Tri-temporal Landsat-OLS change comparison for Seoul. Note multiphase reclamation along coast and post 2001 overglow over water.

(Fig. 5). Near-anniversary Landsat images, acquired during the
growing season clearly distinguish standing crops form urban
development. In most areas urban development takes the form
of substrate–shadow mixtures with periodic 2D texture resulting
from streets and building shadow. The tri-temporal fraction map
shows this expansion as discrete transitions from vegetation to
substrate–shadow mixtures. Several long straight thoroughfares
also appear during the later interval. It is important to note that the
agricultural areas surrounding the development are not completely
unlighted. They correspond to low levels of background luminance
(DN ∼5–10) extending well beyond the overglow of the brighter
areas.
Tri-temporal night lights reveal a more complex multiphase
growth extending outward from a corridor west of Kuala Lumpur
 C. Small, C.D. Elvidge / International Journal of Applied Earth Observation and Geoinformation 22 (2013) 40–52
Fig. 8. Growth in number and area of lights in Asia from 1992 to 2009. Different brightness thresholds show similar number-area trajectories for India, China and all of
southeastern Asia. Insets show growth in cumulative lighted area for China and India. Similar trajectories for different brightness thresholds, countries and spatial scales
suggest consistency in nucleation and growth processes on decadal time scales.
(Fig. 6). In this case, the tri temporal substrate map shows a less
spatially continuous pattern of development with patches of 1990s
change interspersed with patches of 2000s change and large areas
of unchanging vegetation. Close comparison of the tri-temporal
substrate and light composites reveals a disparity between the
area extents of land cover and lighting change. This shows that
brightness changes are often larger in spatial extent than the cor-
responding land cover change. The implication is that changes in
developed areas may overestimate changes in developed area at
fine spatial scales and high brightness levels. For this reason, we
do not generally assume that they correspond to equal changes in
developed area.
Tri-temporal changes in brightness and extent of night lights
in the vicinity of Seoul reveal a very different scenario from the
previous two examples (Fig. 7). While the magnitude of brighten-
ing around Seoul is comparable to New Delhi and Kuala Lumpur,
the spatial patterns shows more infilling and brightening of devel-
oped areas than outward growth. The tri-temporal substrate map
shows most areas unchanged with some areas of reclamation along
the coast. The tri-temporal lights also show a pronounced increase
in overglow over water post 2006 suggesting brightening of the
developed coastal areas. Another limitation of night light is illus-
trated by the area of infilling development near the top center
of the tri-temoral substrate map. Prior to this development the
overglow from the surrounding areas was already saturated so the
infilling development is not captured by the night light. Interstitial
brightening, not corresponding to land cover change suggests that
brightening may occur as a result of increased outdoor lighting in
previously developed areas.
These Landsat-OLS comparisons, along with those previously
published (Small et al., 2011), illustrate two important caveats to
the interpretation of night light imagery – particularly changes in
brightness:
49
1) Changes in brightness correspond to changes in luminance –
not necessarily changes in developed area. Because overglow
is proportional to brightness (Small et al., 2005), increases in
brightness are expected to cause increases in overglow in sur-
rounding areas. Therefore, temporal changes in lighted area may
result from increases in developed area or increases in intensity
of development.
2) Most area of low level luminance (<∼15 DN) does not
correspond to overglow from large bright cities. Large
areas of low level luminance extend over agricultural areas
with numerous small settlements but no large cities (see
www.LDEO.columbia.edu/∼small/DayNight for example com-
parisons). For this reason, changes in brightness should be
considered in the context of the background luminance level
rather than absolute change alone. Therefore the projection
filtered change map should always be used as a reconnais-
sance tool in conjunction with the corresponding calibrated
tri-temporal change map.
6. Comparison of changes in number and extent of lights
Spatiotemporal change in luminance takes two complementary
forms: change in the number of spatially distinct lighted areas
and change in total lighted area above or below some bright-
ness threshold. Together, these result in changes in the total
sum of lights (SoL). These metrics of change represent three dif-
ferent types of change in anthropogenic land cover: nucleation,
aggregation and intensification. Fig. 8 shows the simultaneous
growth of lighted area and number of lights for China, India
and southeastern Asia between 1992 and 2009. These trajec-
tories are shown for three brightness thresholds spanning a
critical percolation transition where the spatial connectivity, and
therefore size distribution, undergoes explosive percolation and
50 C. Small, C.D. Elvidge / International Journal of Applied Earth Observation and Geoinformation 22 (2013) 40–52

Fig. 9. Rank-size distributions in time and space. Each rank-size plot shows the size distribution of all spatially contiguous segments of pixels brighter than11 DN. All slopes
but one lie within 0.15 of −1.0 (thin blue lines). Overall, Asia shows stable growth (slope) but India and China both increase slope from 1992 to 2009 as growth increases
spatial connectedness and larger agglomerations. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of the article.)

becomes superconnected spatially (Small et al., 2011). Despite
different brightness thresholds, spatial scales and geographies,
the area-number trajectories (slopes) for China, India and all of
southeastern Asia are very similar. This suggests a remarkable con-
sistency in nucleation and growth processes on decadal time scales
– despite the very different sizes and growth histories of China and
India (Fig. 8, insets).
In spite of the similarities in growth trajectories, there are some
important differences between China, India and the whole of south-
eastern Asia. Even though India and China have similar areas above
the low brightness threshold (5 DN), China has nearly twice the
area of brighter (DN > 12) lights as India. In 2009, India and China
account for 2 × 105 and 4 × 105 (respectively) of southeastern Asia’s
∼1 × 106 km2 of area brighter than 12 DN. The area-number tra-
jectories for Asia have similar slopes to those of India and China
but are displaced to greater numbers of lights for comparable
areas of light. This reflects the large number of smaller isolated
lights on the archipelagos and coastlines of southeastern Asia com-
pared to the more clustered inland lighted areas in India and
China.
Interestingly, some areas in India appear to have lost consider-
able number and area of lights around 2005. The decreases have
little effect on the sum of lights suggesting that they are mostly
dimmer lights that do not contribute as much to the total sum
of light as they do to the number and area of lights. This appar-
ent change in background luminance, visible in large, spatially
continuous regions of India (Fig. 1), could result from either anthro-
pogenic, atmospheric, imaging or compositing processes. At this
point we lack sufficient information to determine how each process
contributes to the apparent decrease but we point out that
both F15 and F16 composites show significant decreases for
2005 and 2006 in spatially contiguous areas of India – but not
everywhere.
Both area and number of lights well as sum of lights are aggre-
gate measures that reflect changes in the overall spatial distribution
of lighted area. A more complete representation is given by the
rank-size distribution of spatially contiguous lighted segments
and how this distribution changes in space and time. The rank-
size distribution shows all spatially distinct lights sorted by area.
Fig. 9 shows rank-size distributions of lighted areas brighter than
DN = 11 for India, China and all of southeastern Asia. All of the
distributions are strongly linear over four orders of magnitude, sug-
gesting a power law or similar heavy tailed distribution. Similar
distributions are also obtained for DN thresholds of 9 and 6
– but with correspondingly larger areas. The slope and lin-
earity of the distributions do not change appreciably over
this range of thresholds suggesting stability of the underlying
distributions.
 C. Small, C.D. Elvidge / International Journal of Applied Earth Observation and Geoinformation 22 (2013) 40–52
Zipf’s Law predicts a power law rank-size distribution with a
slope of −1 for city size distributions as measured by popula-
tion (Zipf, 1942, 1949), although the linearity and slope of country
level and aggregated population distributions varies considerably
(Nitsch, 2005; Soo, 2005). While the consistency of Zipf’s Law has
attracted sustained interest for several decades, and has been the
basis for a multitude of models (e.g. Berry and Garrison, 1958;
Gabaix, 1999; Lotka, 1941; Pumain, 2006), the varying degree
and extent of agreement with observation impede consensus on
either the universality of the law (Gan et al., 2006; Krugman,
1996; Soo, 2005) or its underlying causes (Batty, 2006; Lotka,
1941; Pumain, 2004, 2006). The assertion of a universal power
law for city size (as represented by population) is controversial
because the estimates of linearity and slope of the power law
rank-size distribution vary over time and among countries (Gabaix
et al., 2004; Nitsch, 2005; Pumain and Moriconi-Ebrard, 1997; Soo,
2005).
Despite mixed results obtained from census enumerations, the
connectivity and size distribution of lighted areas in Asia seems
to be consistent with Zipf’s Law in both space and time. The slope
of all but one of the distributions for brightness thresholds of 6, 9
and 12 DN are within 0.15 of −1 (Fig. 9). This is consistent with
recent global and continental scale analyses of area distributions
of both population density and night light brightness which are
well fit by power laws with slopes near −1 for a wide range of
thresholds (Small et al., 2011). As in the global and continental scale
analyses, the largest spatially contiguous segments in southeastern
Asia are much larger than the individual cities. At scales of 104
to 105 km2 , these large agglomerations correspond to enormous
spatial networks of interconnected cities and smaller settlements
with interstitial areas of lower population or brightness. Yet the size
distributions still conform remarkably well to the power law rank-
size distribution predicted by Zipf’s Law for individual settlements
defined by population.
The temporal consistency of rank-size slope for India, China and
southeastern Asia through time suggests relatively stable growth
throughout the size range over nearly two decades – albeit at very
different rates. The rank-size distribution for India falls proportion-
ately at all sizes in 2005 as expected. Together with the consistent
trajectory of number area plots in Fig. 8, this suggests that aggre-
gate growth processes seem to be quite robust across a range of
areas, geographies and thresholds through time. In spite of this
consistency, some temporal increases in the rank-size distribution
suggest that spatial aggregation is exceeding nucleation in India
and China – but not overall. This apparent increase is analyzed in
greater detail in a separate study.
7. Conclusions and implications
Decadal changes in night light brightness, identified by EOF
analysis, form coherent spatial patterns at local and regional scales.
In areas where sufficient Landsat coverage exists these changes
in brightness correspond to areas where land cover change sug-
gests expansion of anthropogenic land use. The implication is that
multi-temporal night lights can provide a global proxy for lighted
anthropogenic development.
Over the past two decades, consistency in the increase in night
lights in Asia suggest consistency in the underlying processes of
growth and development. Rank-size distributions of spatially con-
tiguous night lights in India, China and throughout southeastern
Asia show strongly linear trends with slopes near −1 for a range
of brightness thresholds. While slopes persist near −1 through
time for southeastern Asia, increasing sub-unity slopes for India
and China suggest that development in these countries is result-
ing in increased connectivity of the larger spatial networks of
51
development at the expense of nucleation of smaller isolated set-
tlements. Taken together with the relatively stable near-unity
slopes of rank-size distributions for southeastern Asia these results
suggest considerable spatial variations in the growth processes
contributing to an apparently stable distribution at continental
scales. The implication is that the logarithmically uniform distribu-
tion observed by Zipf and others may be an emergent phenomenon
arising from spatial aggregation of non-uniform distributions at
finer spatial scales.
Acknowledgment
This work was funded by the United States National Institutes
of Child Health and Development [award R21 HD054846] to CS.
References
Adams, J.B., Smith, M.O., Johnson, P.E., 1986. Spectral mixture modeling; a new anal-
ysis of rock and soil types at the Viking Lander 1 site. Journal of Geophysical
Research 91, 8098–8122.
Batty, M., 2006. Heirarchy in cities and city systems. In: Pumain, D. (Ed.), Hierarchy
in Natural and Social Sciences. Kluwer, Dordrecht, Germany, pp. 143–168.
Baugh, K., Elvidge, C., Ghosh, T., Ziskin, D., 2009. Development of a 2009 stable lights
product using DMSP-OLS data. In: Elvidge, C. (Ed.), Proceedings of the 30th Asia-
Pacific Advanced Network Meeting. Hanoi, Vietnam, pp. 114–130.
Berry, B.J.L., Garrison, W.L., 1958. Alternate explanations of urban rank-size relation-
ships. Annals of the Association of American Geographers 1, 83–91.
Bretherton, C.S., Smith, C., Wallace, J.M., 1992. An intercomparison of methods for
finding coupled patterns in climate data. Journal of Climate 5, 541–560.
Briggs, D.J., Gulliver, J., Fecht, D., Vienneau, D.M., 2007. Dasymetric modeling of
small-area population distribution using land cover and light emissions data.
Remote Sensing of Environment 108, 451–466.
Byrne, G.F., Crapper, P.F., Mayo, K.K., 1980. Monitoring land-cover change by prin-
cipal component analysis of mulittemporal Landsat data. Remote Sensing of
Environment 10, 175–184.
Crist, E.P., Cicone, R.C., 1984. A physically based transformation of thematic map-
per data – the TM tasseled cap. I.E.E.E. Transactions on Geoscience and Remote
Sensing GE-22, 256–263.
Doll, C.N.H., Muller, J.-P., Morley, J.G., 2006. Mapping regional economic activity from
night-time light satellite imagery. Ecological Economics 57, 75–92.
Eastman, J.R., Fulk, M., 1993. Long sequence time series evaluation using standard-
ized principal components. Photogrammetric Engineering and Remote Sensing
59, 1307–1312.
Elvidge, C.D., Baugh, K.E., Kihn, E.A., Kroehl, H.W., Davis, E.R., 1997. Mapping city
lights with nighttime data from the DMSP operational linescan system. Pho-
togrammetric Engineering and Remote Sensing 63, 727–734.
Elvidge, C.D., Tuttle, B.T., Sutton, P.C., Baugh, K.E., Howard, A.T., Milesi, C., Bhaduri,
B.L., Nemani, R., 2007. Global distribution and density of constructed impervious
surfaces. Sensors 7, 1962–1979.
Elvidge, C.D., Ziskin, D., Baugh, K.E., Tuttle, B.T., Ghosh, T., 2009. A fifteen year record
of global natural gas flaring derived from satellite data. Energies 2, 595–622.
Elvidge et al. Remote Sensing of Environment, submitted for publication,
http://www.mdpi.com/journal/remotesensing.
Fung, T., LeDrew, E., 1987. Application of principal components analysis to change
detection. Photogrammetric Engineering and Remote Sensing 53, 1649–1658.
Gabaix, X., 1999. Zipf’s law for cities: an explanation. Quarterly Journal of Economics
114, 739–767.
Gabaix, X., Ioannides, Y.M., Henderson, J.V., Jacques-FranÁois, T., 2004. The Evolution
of City Size Distributions. Handbook of Regional and Urban Economics. Elsevier,
2341–2378,Ch. 53.
Gan, L., Li, D., Song, S., 2006. Is the Zipf law spurious in explaining city-size distribu-
tions? Economics Letters 92, 256–262.
Ghosh, T., Powell, R., Elvidge, C.D., Baugh, K.E., Sutton, P.C., Anderson, S., 2010. Shed-
ding light on the global distribution of economic activity. The Open Geography
Journal 3, 148–161.
Green, A.A., Berman, M., Switzer, P., Craig, M.D., 1988. A transformation for ordering
mutispecral data in terms of image quality with implications for noise removal.
I.E.E.E. Transactions on Geoscience and Remote Sensing 26, 65–74.
Henderson, J.V., Storeygard, A., Weil, D.N., 2009. Measuring economic growth from
outer space. In: National Bureau of Economic Research Working Paper, NBER,
Cambridge, MA.
Johnson, P.E., Smith, M.O., Adams, J.B., 1985. Quantitative analysis of planetary
reflectance spectra with principal components analysis. Journal of Geophysical
Research 90, C805–C810.
Kauth, R.J., Thomas, G.S., 1976. The tasseled cap – a graphic description of the
spectral–temporal development of agricultural crops as seen by Landsat. In:
Proceedings of the Symposium on Machine Porcessing of Remotely Sensed Data,
Purdue University, West Lafayette, Indiana, pp. 4B41–44B51.
Krugman, P., 1996. Confronting the mystery of urban hierarchy. Journal of the
Japanese and International Economies 10, 399–418.
52 C. Small, C.D. Elvidge / International Journal of Applied Earth Observation and Geoinformation 22 (2013) 40–52
Lee, J.B., Woodyatt, A.S., Berman, M., 1990. Enhancement of high specral resoltuion
remote sensing data by a noise-adjusted principal components tranform. I.E.E.E.
Transactions on Geoscience and Remote Sensing 28, 295–304.
Lotka, A., 1941. The law of urban concentration. Science 94, 164.
Nitsch, V., 2005. Zipf zipped. Journal of Urban Economics 57, 86–100.
Preisendorffer, R.W., 1988. Principal Component Analysis in Meteorology and
Oceanography. Elsevier, Amsterdam.
Pumain, D., 2004. Scaling Laws and Urban Systems. Santa Fe Institute, Santa Fe, New
Mexico, p. 26.
Pumain, D., 2006. Alternative explanations of heirarchical differentiation in urban
systems. In: Pumain, D. (Ed.), Hierarchy in Natural and Social Sciences. Kluwer,
Dordrecht, Germany, pp. 169–222.
Pumain, D., Moriconi-Ebrard, F., 1997. City size distributions and metropolisation.
Geojournal 43, 307–314.
Richards, J.A., 1984. Thematic mapping from multitemporal image data using the
principal components transformation. Remote Sensing of Environment 16,
35–46.
Singh, A., Harrison, A., 1985. Standardized principal components. International Jour-
nal of Remote Sensing 6, 883–896.
Small, C., 2004. The Landsat ETM+ spectral mixing space. Remote Sensing of Envi-
ronment 93, 1–17.
Small, C., 2005. Global analysis of urban reflectance. International Journal of Remote
Sensing 26, 661–681.
Small, C., under review. Spatio-temporal dimensionality and time-space character-
ization of multitemporal imagery. Remote Sensing of Environment.
Small, C., Elvidge, C.D., Balk, D., Montgomery, M., 2011. Spatial scaling of stable night
lights. Remote Sensing of Environment 115, 269–280.
Small, C., Pozzi, F., Elvidge, C.D., 2005. Spatial analysis of global urban
extent from DMSP-OLS night lights. Remote Sensing of Environment 96,
277–291.
Smith, M.O., Johnson, P.E., Adams, J.B., 1985. Quantitative determination of min-
eral types and abundances from reflectance spectra using principal component
analysis. Journal of Geophysical Research 90, 792–804.
Soo, K.T., 2005. Zipf’s Law for cities: a cross-country investigation. Regional Science
and Urban Economics 35, 239–263.
Storch, H.V., Zwiers, F.W., 1999. Statistical Analysis in Climate Research. Cambridge
University Press.
Sutton, P., Roberts, D., Elvidge, C., Baugh, K., 2001. Census from Heaven: an estimate
of the global human population using night-time satellite imagery. International
Journal of Remote Sensing 22, 3061–3076.
Townshend, J.R.G., Goff, T.E., Tucker, C.J., 1985. Multitemporal dimensionality of
images of normalized difference vegetation index at continental scales. I.E.E.E.
Transactions on Geoscience and Remote Sensing GE-23, 888–895.
von Storch, H., Zwiers, F.W., 1999. Statistical Analysis in Climate Research. Cam-
bridge University Press, Cambridge, UK.
Zipf, G.K., 1942. The unity of nature, least-action, and natural social science. Sociom-
etry 5, 48–62.
Zipf, G.K., 1949. Human Behavior and the Principle of Least-Effort. Addison
Wesley.